Custody chains show who currently owns autonomous work, why it moved, and what authority moved with it.
Which system, worker, rule, or human judgment created the work item?
// AGENT OPS · AI-NATIVE ORG DESIGN · June 8, 2026
Machine Workers Need Custody Chains Before Work Crosses Teams
When agent work crosses a team boundary, ownership cannot travel as vibes.
Machine workers are good at connecting systems. Companies are bad at noticing when connected systems erase accountability.
AI-native teams get more cross-functional throughput without turning shared work into orphan work.
Autonomous work does not fail only inside a task. It fails between teams.
Most agent-ops design focuses on the worker: the prompt, the tool permission, the retry policy, the human approval step. That is necessary, but incomplete. The brittle moment is often the transfer.
A prospect signal becomes a sales action. A support complaint becomes a product escalation. A billing anomaly becomes a customer-success risk. A campaign insight becomes a RevOps rule change. The agent did not just complete work; it changed the custody of the work.
If the handoff lacks a custody chain, the organization gets motion without accountability. The task appears in a new place, but the old context, risk, and decision rights do not reliably arrive with it.
The custody chain has five links.
What evidence justified moving the work across a team or tool boundary?
Who is accountable for the next decision after the handoff lands?
What may the receiving worker or team do without new approval?
When does the handoff become stale, rejected, or forced back to review?
The dangerous version is a correct handoff with missing custody.
Bad agent work is easy to criticize. Correct agent work with weak ownership is harder because everyone agrees the signal mattered.
Example: a lifecycle worker identifies a renewal account with rising support frustration and creates an executive escalation. The signal is real. The risk is real. But if the escalation lands without a current owner, evidence trail, and authority boundary, it becomes a Slack-shaped anxiety object. People discuss it. Nobody knows who is allowed to decide.
That is how autonomous systems create management fog. They increase detection before the company has upgraded decision custody.
A grounded GTM example
Without custody
The agent creates urgency and spreads ambiguity.
A deal-intelligence worker notices that a strategic account visited pricing, opened three security docs, and filed a support ticket from a related domain. It creates tasks for sales, success, and support.
Each task looks reasonable. Together, they create overlapping outreach, duplicate research, and three teams wondering who owns the customer narrative.
With custody
The signal enters one accountable lane.
The worker attaches the evidence, assigns sales as temporary custodian for commercial action, limits support to ticket resolution, and gives success read-only context until sales accepts or rejects the account motion.
The work still crosses systems. The authority does not leak everywhere.
What the receipt should carry
Original surface
The CRM view, ticket queue, product event, billing record, campaign result, or agent run that created the signal.
Transfer reason
The specific evidence that made cross-team movement necessary instead of local resolution.
Current custodian
The person, queue, role, or machine worker responsible for the next decision.
Allowed actions
The actions permitted after handoff: draft, enrich, notify, update status, contact customer, change owner, or block.
Return path
Where the work goes if rejected, stale, duplicated, low confidence, or outside the receiving team’s authority.
Custody chains are not bureaucracy. They are the price of parallelism.
Human teams often rely on social memory to manage handoffs. People know who usually owns renewals, strategic accounts, campaign exceptions, and support escalations. Machine workers do not inherit that social fabric automatically.
As agent labor scales, parallel work increases faster than shared understanding. Custody chains make the implicit operating model explicit enough for machines to use and humans to audit.
The goal is not to slow every transfer. The goal is to make cross-boundary work safe enough to move quickly.
If nobody owns the next decision, the agent did not hand off work. It created noise with evidence attached.
Operator checks
Three signs your custody chain is missing
Two teams act on the same agent-created signal because the worker notified everyone instead of assigning one custodian.
The receiving team sees the task but not the reason, source, confidence, or boundary that created it.
The task is discussed across channels, but nobody can say who can approve, reject, expire, or return it.
Install custody at the handoff layer.
List where autonomous work crosses teams: sales to success, support to product, RevOps to finance, marketing to SDR, customer data to lifecycle.
Require source, reason, owner, authority, and expiry on every cross-boundary task.
Review tasks that changed systems but never received an owner decision.
Default to one custodian and informed observers, not five teams with equal implied responsibility.
Concrete takeaway for this week
Pick one agent workflow that creates cross-functional tasks and add custody fields before you add more automation.
- Name the custodian: the role or queue responsible for the next decision.
- Name the evidence: the source signal that justified the handoff.
- Name the authority: what the receiving side may do without fresh approval.
- Name the expiry: when the handoff becomes stale or must return to review.
- Fix: stop letting agents broadcast responsibility when the work needs one owner.
The blunt rule: every machine-generated handoff should answer who owns this now, why, and until when.